niessner/Opt

Proposition / discussion about introducing hashing data structure support

Algomorph opened this issue · 0 comments

I'd like to propose introducing support for indexing / data structures (for instance, spatial hashing) that separates data into blocks, and where each block can be processed by a separate thread.

If I find that the task is feasible (extending Opt and, potentially, Terra itself doesn't require a major re-engineering effort), I am willing to take the lead on this. That being said, I am a PhD student, and am open for collaboration if this work can lead to a publication (this would involve demonstrating advantages this would yield for different computer vision/graphics tasks).

For the authors and current maintainers, I have some questions:

  1. What potential changes do you foresee to best realize support for extendable data structures (and a spatial hashing implementation) in Opt? Would this warrant a new data structure / class, such as the Array or SampledArray? Or would this involve a new access method, such as a Stencil/Graph? What difficulties / complications do you foresee? Would modifications to Terra be inevitable, and, if so, what kind of modifications?

  2. Perhaps you or someone you know would be willing to join the effort ( or assist with future questions, being domain experts in Opt and/or Terra)? Since I would be targeting a publication, I am open to co-authorship provided reasonable effort/assistance.